AI21's forward-deployed researchers post-trained NVIDIA's open Nemotron into custom models for Gong, Deel and Alice in three weeks

AI21 Labs published a post on October 15 describing three customer projects run by its Forward-Deployed Researchers (FDRs), a research-heavy cousin of the forward deployed engineer. Over roughly three weeks, the team post-trained NVIDIA's newly released open model Nemotron 3.5 Super VL with Gong, Deel and Alice, building each customer's training data, training recipe and custom evaluations as part of what AI21 sells as a Lab-as-a-Service offering.

AI21 reports that Gong's own evaluations found the tuned model matched Claude Opus 5.5 on speed and accuracy for its CRM Data Extractor at a fraction of the cost. On Deel's payroll questions, accuracy reached 72% on an 8-country benchmark from Emet by Deel, 8 points above Opus 5.5 at 33x lower cost per query. With Alice's security RL environments, resilience to indirect prompt injection in Slack and GitHub tasks improved by 41 points without hurting task performance. These are the company's own figures.

For FDEs, this shows the role moving deeper into the model stack. Customer-embedded work used to mean wiring models into workflows; here it means assembling datasets, choosing techniques like SFT, DPO and GRPO, and building eval gates for high-risk outputs on a customer's own use case. As more enterprises try to own specialized models instead of renting general ones, expect more "forward-deployed researcher" and post-training-focused FDE roles at model labs and platforms.